Syntactic/Semantic Structures for Textual Entailment Recognition

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 1020–1028, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Syntactic/Semantic Structures for Textual Entailment Recognition Yashar Mehdad FBK-IRST, DISI University of Trento Povo (TN) - Italy mehdad@fbk.eu Alessandro Moschitti DISI University of Trento Povo (TN) - Italy moschitti@disi.unitn.it Fabio Massimo Zanzotto DISP University of Rome “Tor Vergata” Roma - Italy zanzotto@info.uniroma2.it Abstract In this paper, we describe an approach based on off-the-shelf parsers and semantic re- sources for the Recognizing Textual Entail- ment (RTE) challenge that can be generally applied to any domain. Syntax is exploited by means of tree kernels whereas lexical se- mantics is derived from heterogeneous re- sources, e.g. WordNet or distributional se- mantics through Wikipedia. The joint syn- tactic/semantic model is realized by means of tree kernels, which can exploit lexical related- ness to match syntactically similar structures, i.e. whose lexical compounds are related. The comparative experiments across different RTE challenges and traditional systems show that our approach consistently and meaningfully achieves high accuracy, without requiring any adaptation or tuning. 1 Introduction Recognizing Textual Entailment (RTE) is rather challenging as effectively modeling syntactic and semantic for this task is difficult. Early deep seman- tic models (e.g., (Norvig, 1987)) as well as more re- cent ones (e.g., (Tatu and Moldovan, 2005; Bos and Markert, 2005; Roth and Sammons, 2007)) rely on specific world knowledge encoded in rules for draw- ing decisions. Shallower models exploit matching methods between syntactic/semantic graphs of texts and hypotheses (Haghighi et al., 2005). The match- ing step is carried out after the application of some lexical-syntactic rules that are used to transform the text T or the hypothesis H (Bar-Haim et al., 2009) at surface form level. For all these methods, the ef- fective use of syntactic and semantic information de- pends on the coverage and the quality of the specific rules. Lexical-syntactic rules can be automatically extracted from plain corpora (e.g., (Lin and Pantel, 2001; Szpektor and Dagan, 2008)) but the quality (also in terms of little noise) and the coverage is low. In contrast, rules written at the semantic level are more accurate but their automatic design is difficult and so they are typically hand-coded for the specific phenomena. In this paper, we propose models for effectively using syntactic and semantic information in RTE, without requiring either large automatic rule acqui- sition or hand-coding. These models exploit lexi- cal similarities to generalize lexical-syntactic rules automatically derived by supervised learning meth- ods. In more detail, syntax is encoded in the form of parse trees whereas similarities are defined by means of WordNet simlilarity measures or Latent Seman- tic Analysis (LSA) applied to Wikipedia or to the British National Corpus (BNC). The joint syntac- tic/semantic model is realized by means of novel tree kernels, which can match subtrees whose leaves are lexically similar (so not just identical). To assess the benefit of our approach, we carried out comparative experiments with previous work: especially with the …

Content truncated to 500 words